Predicting past and future avalanche danger in northern Norway with machine-learning models

Snow avalanches are one of the most impactful natural hazards in mountainous areas. Avalanche characteristics are likely to change in a changing climate, especially in the Arctic where changes are more rapid, posing a severe challenge for local adaptation. Here we train machine-learning (ML) models to predict avalanche danger in northern Norway and then apply these models to dynamical downscalings of future climate projections.

We utilise regional expert avalanche-danger level assessments differentiating two different avalanche problems: wind slab, and wet (loose and slab combined). The ML models are trained on the 3-km Norwegian reanalysis (NORA3) to estimate the linkage between avalanche danger in the Troms region of Norway and local meteorological conditions. For the future climate simulations, we employ the Nordic Convection Permitting Climate Projections (NorCP), providing a 3-km dynamical downscaling of the Representative Concentration Pathway (RCP) scenarios performed with two global climate models.

To obtain a rough estimate of the trend of avalanche danger, the European Avalanche Danger Services (EAWS) 5-level avalanche danger scale is changed into a binary setup with levels 1 and 2 aggregated to 0 and levels 3, 4, and 5 to 1. The overall accuracy of the ML model for the wind slab problem is about 80 % and considerably higher than for the wet problem with about 67 %. This indicates that while the wind slab problem is to a high degree determined by the recent weather, this is less so for the wet problem. Including information from sophisticated snowpack modelling in the training data may thus increase the prediction accuracy.

By applying the ML models in a hindcast setting to the whole NORA3 record (1970-2024), we find a correlation between avalanche danger and a well-known climate mode, namely the Arctic Oscillation (AO). Given recent advances in model skill in representing the AO, this has potential implications for the seasonal predictability of avalanche danger in northern Norway.

Moreover, by applying the ML model to the NorCP simulations (2040-2060 and 2080-2100), the results differ per avalanche problem: While the wind slab avalanche danger declines in all scenarios, the wet avalanche danger remains relatively constant and even increases in some cases. The former appears to be related to decreasing snowfall and wind speed, while the latter is likely connected to increasing temperatures and rain.